Overview
- Nvidia publicly showed a Rubin Ultra tray with four compute chiplets and 1TB of HBM4E memory, billed as the first AI accelerator to reach a terabyte of on‑package memory, with availability targeted for 2027.
- Rubin Ultra will deploy in new Kyber racks using vertical, liquid‑cooled trays to fit 144 GPU packages per NVL144 system, which Nvidia says will deliver at least four times the performance of current NVL72 designs.
- Kyber platforms introduce an NVLink Gen‑7 switch to increase GPU counts at consistent link speed and add Nvidia’s CX9‑1600G Ethernet processor to accelerate scale‑out networking.
- Nvidia’s roadmap sets 2027 Rubin Ultra pairings with Groq’s LP35 LPU supporting the NVFP4 data format, followed in 2028 by the Feynman generation with 3D stacking, custom HBM, Rosa CPUs, and LP40.
- Nvidia plans NVLink switches with co‑packaged optics in 2028 to extend rack‑scale systems to as many as 576 or 1,152 GPU packages, while some implementation specifics such as final silicon and tape‑out timing remain undisclosed.